What hardware do you need to run Arete-27B?

Arete-27B at Q4_K_M is 17 GB of weights, with a context ceiling of 256k tokens. Nous Research fine-tuned Alibaba's Qwen3.8-27B to change how it answers on politically sensitive subjects and left the rest alone: supervised fine-tuning on 52.8 million tokens of text over two epochs, no reinforcement learning, no change to any layer, and the vision encoder frozen, with another open model, Inkling, as the teacher. On Nous's own 803-prompt evaluation, judged by Grok 4.7, factual recall rose from 2.32 to 2.98 out of 5 and the share of answers taking the Chinese state's framing fell from 22% to 0.9%. Those are Nous's own figures rather than an independent test, and nobody has placed this model on the index the rest of this site is ranked by. The checkpoint also ships without the base model's multi-token-prediction head, so it carries slightly fewer parameters than Qwen3.8 27B even though the Q4_K_M file priced here is the larger of the two. On the cheapest machine that runs it, a Framework Desktop, 32GB, it generates 9.9 tok/s at 32k of context, worked out from memory bandwidth.

How good is Arete-27B, really?

It has not been placed on the intelligence index yet. See the whole table.

Head to head with the models that need much the same memory at 32k of context: vs Qwen3.8 27B, vs Qwen3.6 27B, vs Gemma 3 27B it.

What Arete-27B costs either way

Nobody rents Arete-27B by the token. The closest hosted match, Qwen3.8 27B, costs $0.039 per million input tokens and $1.25 per million output (OpenRouter, cheapest active endpoint, checked 2026-11-10). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill. What a million tokens costs each way puts the two prices side by side.

Machines that run Arete-27B

Every machine here runs it, but not to the same length. The Mac mini M6, 32GB stops at 32k; the DGX Spark GB10 Grace Blackwell, 128GB takes it to 256k, which is this model's own ceiling and the longest the calculator offers. The weights are the same size on every machine; what differs is the memory left for the key-value cache, which grows with every token you keep.

The table is the cheapest machine in each family, not the only one that runs it: 44 of the 56 machines on this site hold it at 32k. Within a family only memory decides, so that is every Mac Studio and Strix Halo box at any size, every NVIDIA card with 24GB or more, every Mac mini and MacBook Pro with 32GB or more, the Radeon AI PRO R9700, 32GB and the DGX Spark, 128GB. One table sets every machine against the others: price, the memory its GPU can use, what it holds and how long it takes to pay for itself.

One machine per family, cheapest first. Speeds are measured where a public benchmark exists and estimated from memory bandwidth otherwise; the calculator says which for any configuration. The longest context is the longest setting the calculator offers that the machine still holds this model at, cache included, and no machine is shown taking it past its own 256k limit. Each one opens the calculator on that machine at that length.

For scale, the calculator starts from 80 tok/s for a hosted API and times a local machine against it. Nothing above reaches it, and the quickest is 64 tok/s.

The cheapest machine that runs it is sold with 32 GB. There is a page on what 32 GB runs, machine by machine: it names the 6 machines here sold at that size, and the 37 current models the roomiest of them holds at 32k, this one among them.

It fits in less memory than that: the smallest size here that runs it at 32k is 24 GB, on the GeForce RTX 3090, 24GB at $1,499, card only, which hands a model 23 GB and is no longer sold.

Run Arete-27B at 256k on the DGX Spark GB10 Grace Blackwell, 128GB

How much memory does Arete-27B need at 256k?

Every figure above is priced at 32k of context. At 256k, which is this model's own ceiling and the longest the calculator offers, Arete-27B needs 34.4 GB rather than 19.3 GB. The weights are the same either way; the key-value cache goes from 2.1 GB to 17.2 GB.

30 of the 37 machines here still sold at a published price hold Arete-27B at 32k and 24 hold it at 256k. The cheapest still sold goes from the Framework Desktop, 32GB at $1,269 to the Corsair AI Workstation 300, 64GB at $1,700.

On the Corsair AI Workstation 300, 64GB it runs 9.9 tok/s estimated at 32k and 5.6 tok/s estimated at 256k. Every token generated reads the whole cache back, so a longer window is slower as well as bigger.

The specifics

Parameters

27.4B

Quantisation

Q4_K_M

Weights on disk

17 GB

Images

It reads images as well as text, and the weights above are the language model. Feeding it pictures loads a separate 0.9 GB vision projector, mmproj-NousResearch_Arete-27B-f16.gguf, on top of every figure on this page.

KV cache

2.1 GB at 32k context — Hybrid: 48 linear-attention layers; only 16 full-attention layers hold a growing KV cache. The fine-tuning changed no layer of Qwen3.8 27B, so the geometry here is that model's. How weights and cache add up.

Maximum context

256k tokens (the same ceiling as the model it was fine-tuned from; Nous trained it at 64k)

Licence

Apache 2.0

Sources

NousResearch/Arete-27B, NousResearch/Arete-27B (config), bartowski/NousResearch_Arete-27B-GGUF, openrouter.ai